Hybrid Bayesian Neural Networks for Li-ion Battery Prognosis

Researchers have proposed a hybrid Bayesian physics-informed neural network framework for Li-ion battery prognosis, addressing challenges such as computational efficiency, partially characterized first-principle models, and unstructured datasets. The framework integrates numerical integration of governing equations using recurrent neural networks, compensates for model-form uncertainty with data-driven nodes, and incorporates variational Bayesian nodes to account for data uncertainty. The approach allows for accurate prediction of battery end of discharge while considering the effect of battery aging, inter-battery variability, and temperature effects. The hybrid model can handle diverse sources of information and offers benefits such as model updating without decommissioning the battery, handling battery-to-battery variation, and incorporating fleet-wide data for battery degradation modeling. Experimental data from the NASA Prognostics Center of Excellence Data Repository is used to demonstrate the effectiveness of the framework.